Subway ridership dynamics
The dynamics of COVID-19 outbreak, intervention measures, and subway ridership are demonstrated in Figs. 2 and 3. In detail, ridership is equivalent to the sum of all daily subway entries in each city and represents individuals’ behavioral choice in response to COVID-19.
In Seoul, ridership slightly declined after the Orange-alert level was alarmed on January 27th and then further declined after the first COVID-19-induced death occurred on February 20th. This drastic decline continued through mass outbreaks in Daegu and Kyung-Buk provinces, where cases reached peak points on February 29th. Thereafter, ridership began to recover until it flattened starting the second week of May. However, due to a seasonal declining pattern also shown in 2019, the ridership patterns appear to reflect previous counts and maintain an existing trend. Then, as cases per million people began to surpass that of the February 29th peak in mid-August, social distancing guidelines were raised to Level 2 starting August 16th. During this period, ridership began to decline like that of late-February, but the rate of decline was less severe. However, ridership rebounded in early-September when the city heightened the social distancing guidelines to Level 2.5 after outbreak that started mid-August persisted. After the rebound, ridership began to recover through late September.
Seoul’s dynamics trend indicates that subway ridership acutely dropped after specific events and surged after interventions were implemented. Also, ridership recovery in both late-February and early-September occurred regardless of the location of the outbreaks while the rate of recovery was faster in the latter period than the former. These observations suggest that national outbreak events affect Seoul’s ridership, and the magnitude of that impact could be influenced by the timing of the outbreak. Conclusively, individual behavior could be more affected by the outbreak timing than by the location of the outbreaks.
In NYC, similar to that of Seoul, subway ridership pattern between 2019 and 2020 is similar until the first COVID-19 case was confirmed on February 29th. Thereafter, ridership began to decline and declined further starting March 8th, when NYC began to issue mitigation interventions. Ridership further declined after the state announced the “New York on PAUSE” (PAUSE) order, but began to bounce back starting April 15th, when face masks or coverings were mandated in public places. As the recovery trend continued, the rollout of Reopening Plan Phase One on June 8th accentuated the ridership increase. This incline continued through Phase Two and began to stagnate after Phase Three was introduced on July 5th. However, like Seoul’s ridership pattern, NYC’s appears to reflect previous year’s trend considering seasonal factors. A minutely inclining pattern started in early September, when COVID-19 cases began to remain under 250 people per day, movement and indoor activity restrictions were eased, schools and universities gradually reopened, and other reopening guidelines took place.
NYC’s dynamics trend indicates that NYC’s subway ridership was largely affected by the outbreak in early stage of the pandemic. Generally, ridership appeared to decline after restriction measures were introduced or intensified while it increased after public health guidelines, such as face mask or covering mandates, and reopening plan phases were implemented. A more microsopic observation of NYC’s ridership reveals that, despite a national increase in COVID-19 cases, NYC’s subway ridership gradually recovers after the sharp decline in the initial stage of the pandemic. This observation suggests that NYC’s ridership is significantly affected by national outbreak events, but by government mitigation strategies. As a result, NYC residents’ behavioral response to COVID-19 is more affected by government interventions than by the outbreak pattern.
Comparing the dynamics of both cities, NYC’s ridership decreases more significantly than Seoul’s at the onset of the pandemic, mid-February in Seoul and early-March in NYC. The difference in these trends can be attributed to the difference in their disease response strategy per the cities’ differing scale and gravity of the pandemic’s impact. Whereas Seoul facilitated multiple softer guidelines like social distancing, quarantine, and contact tracing, NYC enforced more stringent rules through PAUSE. PAUSE required all non-essential businesses to close, canceled or postponed non-essential meetings, and restricted individuals’ mobility. In contrast, Seoul’s strategy centered on encouraging social distancing and wearing of face masks, regulating indoor dining, and limiting group gatherings but not entirely restricting individuals’ movements.
Other than the intervention level, factors such as personal perception of risk, ability to work remotely, and the feasibility of travel mode shift could be essential factors for the different ridership pattern in Seoul and NYC. Although the two cities’ subway modal share is similar—NYC’s 38.5% (U.S. Census Bureau 2020) and Seoul’s 40.7% (Seoul Metropolitan Government 2020)—their paths for subway ridership recovery following the initial COVID-19 outbreak diverge. In NYC, where ridership recovery occurred at a slower pace, travel through personal vehicle recovered the earliest (Ye et al. 2021; Kamga et al. 2020). Further, when restrictions of subway use were imposed, use of personal vehicles and bike share systems increased (Bian et al. 2021), while use of taxis increased in areas with higher low-income populations and dependency on public transit (Manley et al. 2021). Such travel behaviors and modal shift among New York City residents reveal the various factors affecting shift from subways to alternative travel options during COVID-19.
The two cases suggest that government interventions that restrict individuals’ mobility cause COVID-19 cases to decline. In response to the mass outbreak earlier in the pandemic, NYC shut down the city with austere movement restrictions and gradually reopened the city as the disease outbreak stabilized and eventually declined. In Seoul, however, social distancing and other mitigation strategies were administered but individual mobility was not severely restricted to the level of a shutdowns. As such, even after the mass outbreak in Daegu and Kyung-Buk that led to the pre-second wave period near June 22nd, subway ridership in Seoul did not significantly reduce and continued through the second wave that centered on Seoul starting mid-August.
The differing approaches of NYC and Seoul reveal the limits of shutdown and social distancing measures on mitigating the effects of COVID-19. For instance, although NYC’s shutdown measures effectively curbed the surge of COVID-19 cases, they may have left negative socioeconomic effects on the livelihood of New Yorkers, especially for low-income families and people of color. For instance, almost half of all NYC workers lost employment income at the height of the pandemic (Collyer et al. 2020); however, differences in essential worker status contributed to deepening racial and socioeconomic disparities in COVID-19 transmission. Black and Hispanic workers were more likely to be required to work on-site (Collyer et al. 2020), facing increased risk of transmission in the workplace and via public transit (Dubay et al. 2020; Hawkins 2020). Simultaneously, low-wage workers were less likely to have access to paid leave to stay home even if they had become sick or were exposed to COVID-19 (US Bureau of Labor Statistics 2021), further risking their economic and physical well-being. Similarly, Seoul’s social distancing measures helped slow the spread of COVID-19 at an early stage; however, its long-term effectiveness was disproved in its compact city form. This result implies that different countries’ COVID-19 response strategies could yield place-blind and uniform results, and confirms that OECD’s (2020) recommendation to apply both place-based and people-based approaches to control the pandemic could be effective.
Time-series clustering of Seoul subway stations
Following the dynamics analysis, time-series clustering of the 296 subway stations in Seoul was conducted to formulate six clusters based on time-series ridership pattern (Table 2). Cluster one had 75 stations, cluster two had 72 stations, cluster three had 76 stations, cluster four had 32 stations, cluster five had 28 stations and cluster six had 13 stations (Fig. 4).
In all stations, subway ridership declined following the first confirmed case and outbreak of COVID-19, which is January 24, 2020, for Seoul and January 20, 2020, for South Korea nationally (Fig. 5). After the decline, ridership generally increases in all stations, which signaled some resilience towards bouncing back to pre-COVID-19 levels. However, stations with higher ridership counts exhibited delayed recovery to pre-COVID levels. Also, though the variance between pre- and post-COVID ridership narrows, the order of change in variance among stations does not change. As such, the overall ridership pattern appears to not have changed despite the reduction in total ridership. In particular, more frequented stations exhibited less evident changes in ridership pattern between pre- and post-COVID levels.
Comparing the difference in total subway ridership pre- and post-COVID offers a more detailed picture (Table 3), in which smaller C values signal stronger resiliency.
The clustering results reveal that subway stations with typically higher ridership level underwent larger fluctuation in ridership, and stations with lower ridership level underwent smaller fluctuation in ridership. In the context of individual behaviors related to COVID-19-related risks (Bucsky 2020; Li et al. 2020a; Shamshiripour et al. 2020; Abdullah et al. 2021; Kim et al. 2021), this result suggests a possible close association between subway ridership and travel purpose. For instance, ridership fluctuation at stations linked with non-mandatory trip purposes, like leisure, will likely be non-stationary and ridership fluctuation at stations linked with mandatory trip purposes, like commute for work, will likely be more stationary. In Seoul, ridership at stations linked with mandatory trip purposes will likely be lower since the residences of people with jobs in Seoul are distributed across the greater Seoul Metropolitan Area.
The result of the ridership-based clustering analysis suggests that the impact of COVID-19 varies across stations and reveals that the fluctuation in ridership pre- and post-COVID-19 varies according to the scale of ridership levels. This implication is attributed to the idea that individuals’ perception of and response to COVID-19 risks could be reflected by their trip purpose and personal choice of travel mode, represented as subway ridership level.
Figure 6 illustrates ridership level at each cluster’s medoid station with key pandemic-related events and interventions. Cluster 6 stations have prominently high and mire fluctuating ridership levels over time, which implies a more elastic nature of demand for Cluster 6 stations. This finding also suggests that Cluster 6 stations could act as disease transmission since higher ridership level alludes to greater mobility and possible overcrowding (Hamidi et al. 2020; Bhadra et al. 2021). With smaller degrees in fluctuation, Clusters 4 and 5 stations exhibit similar patterns of high and more fluctuating ridership level. In contrast, Clusters 1, 2, and 3 all display similar patterns of average ridership level with minutely different and relatively small fluctuation range. As shown in Table 4, the mean, standard deviation, and variance all increase as the Cluster numbers increase, which implies that stations with higher daily ridership counts face greater fluctuation in ridership. As such, the study finds that stations with higher ridership levels endure wider fluctuation in ridership following disease spread, implying that contagious diseases like COVID-19 leaves greater impact on stations where overcrowding occurs more frequently.
Understanding the land use mix in Seoul helps contextualizing these findings. Overall, land use mix in station catchment area within 500 m radius consist of, in descending order, residential area, road and transportation facilities, commercial and office area, and public facilities and institutions (Fig. 7). Simply classifying the six clusters into categories based on ridership levels, catchment area of larger-scale stations (Clusters 4, 5, 6) have higher proportion of commercial and office area, industrial and manufacturing area, and road and transportation facilities than that of smaller-scale stations (Clusters 1, 2, 3). In contrast, the catchment area of smaller-scale stations have higher proportion of agricultural and forestry area, public facilities and institutions, and residential area than that of larger-scale stations.
To assess the viability of compact cities against shocks like COVID-19, this study analyzed the association between subway ridership decline and land use mix, a key feature of compact cities. Controlling sociodemographic and station attribute variables, Model 1 suggests that land use mix is associated with the reduction in subway ridership after COVID-19, confirming previous findings that higher degree of diversity in land use positively contributes to subway usage (Choi et al. 2012). This result reverses in the context of COVID-19, during which subway ridership decreased in highly mixed land use areas. As such, land use mix appears to affect the overall volatility of ridership levels.
From the full model, Model 2, statistically significant variables with the highest coefficients are Pempl_accfood (4.343), Proad_transpfac (2.583), Ppublfac_instit (1.554), Pcomm_office (1.256), Landusemix (0.747), Transfer (0.719), Busstop (0.015), Activitydens (0.011). Controlling various land use classifications, the coefficient of the land use mix index (Landusemix) decreases, but its statistical significance is retained as was in Model 1. Model 2’s result for land use mix index demonstrates that other land use classifications that are not controlled in this model exist in compact cities like Seoul; therefore, people’s mobility may be restricted by risks like COVID-19 in areas with diverse land use, notably compact cities.
Of the land use variables, the ratio of road and transportation facilities area to total land area (Proad_transpfac) could stand for accessibility in a catchment area since it supports on-ground modes of travel alternative to subways, such as personal vehicles, taxi, and bikes. Model 2 results show that Proad_transpfac is statistically significant with positive impact on subway ridership decline. In line with other studies (Abdullah et al. 2020; Teixeira and Lopes 2020), this result implies that areas with higher accessibility via non-subway travel modes faced greater decline in subway ridership after the COVID-19 outbreak. On the other hand, the ratio of public facilities and institutions (Ppublfac_instit), also statistically significant and positively associated with ridership decline, appear to reflect the impact of COVID-19 risk management interventions—transitioning to remote learning for schools and limiting access to multi-purpose facilities like sports stadiums—on reducing mobility overall. Similarly, the ratio of commercial area and office space to total land area (Pcomm_office), statistically significant and positively associated with ridership decline, further contextualize the impact of government COVID-19 prevention measures on reducing mobility. For instance, restrictions on accommodation and food services discouraged large gatherings likely reduced traffic in commercial areas. Likewise, transition to remote work following the government’s escalation of the risk alert level to Red likely led to reduced crowding in areas with significant office space.
Other than land use variables, the ratio of jobs in accommodation and food services industry to total jobs (Pempl_accfood) had the largest statistically significant impact on the decline in subway ridership after the COVID-19 outbreak. This result could be credited to the Korean national and Seoul’s municipal COVID-19 risk management interventions, which included social distancing measures and restrictions on operation of restaurants, cafes, and other multi-purpose facilities. More specifically, these restrictions mandated earlier facility closing, limited indoor dining, and shutdown of large-scale facilities. When COVID-19 cases peaked to anomalously high records, these measures were complemented by more austere measures restricting the use of public facilities and gatherings were imposed. Complementing these measures, the public’s risk perception toward multi-purpose facilities seemed to have been high during the pandemic (Kim et al. 2021). As such, the impact of Pempl_accfood on ridership decline could be explained by constrained mobility due to restriction guidelines and the public’s high risk perception of COVID-19.
Activity density (Activitydens) had a relatively smaller coefficient but was statistically significant. This result signifies that subway usage is lowered in areas with high activity density, which corresponds to past findings that activity density is associated with COVID-19 infection and mortality rates (Hamidi et al. 2020). Other sociodemographic variables pertaining age help contextualize whether ridership pattern is associated with working population (P20to64) and retired-age population (P65over) per past studies that identified reduced ridership among retired-age population during COVID-19 (Park and Cho 2021). Model 2 results reveal that these variables are not statistically significant, but their coefficient measures indicate some influence on reduced ridership during the pandemic. In a city with high subway modal share, Seoul’s subway modal share at 41.6% in 2019, the impact of both working and non-working populations on ridership decline suggests that modal shift may be one of individuals’ behavioral changes resulting from the COVID-19 outbreak.
Together, variables for the number of bus stops in catchment area (Busstop) and transferability (Transfer), show modal shift that alludes to connectivity among different travel modes. Both variables are statistically significant and are positively associated with ridership decline, indicating that stations located in area with fluid modal shift between bus and subway experienced ridership decline during the pandemic. This result aligns with other findings that confirm positive impact of intermodal network on subway ridership (Cervero 1996; Lee et al. 2013). Of the two, the impact of Transfer was stronger than that of Busstop, suggesting that shift within same travel mode leaves greater impact on ridership change than shift between different travel modes (Guo and Wilson 2007) (Table 5).